Robot tracking error control method based on neural network model prediction compensation
By using the duration recurrent neural network (GACTRNN) based on gating and adaptive mechanisms in industrial robots, the problem of trajectory tracking accuracy and stability of industrial robots under large load and strong impact conditions is solved. The effect of high-precision trajectory tracking is achieved.
Patent Information
- Application Number
- CN202411238679.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art is difficult to maintain the motion accuracy and system stability of trajectory tracking in industrial robots, especially under large loads, strong impacts, and dynamic load conditions. Traditional control algorithms cannot meet the requirements of high-precision trajectory tracking.
A duration recurrent neural network (GACTRNN) based on gating and adaptive mechanisms is used to build a neural network model. Through the "sequence to sequence" processing method, a neural network containing encoder and decoder is built, and a hyperbolic tangent function is used as Activation function regulates the update/forgotten rate of neurons, and realizes prediction and compensation of time series signals.
Through neural network models, the tracking error of industrial robots is learned and generalized, so as to maintain the motion accuracy and system stability of tracking, and meet the requirements of high-precision tracking of robots.
Smart Images

Figure CN119065252A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of robot tracking error control, and in particular to a robot tracking error control method based on neural network model prediction compensation. Background Art
[0002] Industrial robots are multi-joint manipulators or multi-degree-of-freedom machine devices used in the industrial field. They are the core equipment for the development of the country's manufacturing industry. They can rely on their own power and control capabilities to achieve various industrial processing and manufacturing functions. Industrial robots are widely used in various industrial fields such as electronics, logistics, and chemicals. As key industrial structural parts are developing towards large-scale and complex curved surfaces, the country's key manufacturing industries have put forward urgent demands on industrial robots and their application performance.
[0003] Neural network (NN) is a modern mathematical modeling tool inspired by neuroscience research. It can be used to mathematically model input and output data or identify data patterns. The basic unit of a neural network is a neuron. A single neuron performs certain operations on input data through linear operations and activation functions. Multiple neurons are combined into a neural network layer. The stacking of multiple neural network layers constitutes a complete neural network. By training the neural network with data, the neural network can be enabled to handle specific data tasks.
[0004] Based on the research of brain science and perception science, the present invention uses a neural network variant based on brain-like intelligence to extract periodic component signals from time series signals - a duration recurrent neural network (GACTRNN) based on a gating and adaptive mechanism. According to the current external input, multiple groups of time scale parameters can be adjusted in real time, and the internal state update speed of neurons can be further changed to regulate the memory / forgetting rate, thereby enabling it to have a powerful processing capability for periodic signals.
[0005] This network is a type of recurrent neural network (RNN). According to the length of input data and output data, the application modes of RNN can be divided into four categories: one-to-one, one-to-many, many-to-one and many-to-many. Among them, the many-to-many method is also called the "sequence to sequence" method. It can predict the subsequent sequence based on the current string of input sequences, and the input sequence and output sequence can have different lengths.
[0006] At present, there are few industrial robots in my country with stable performance and high reliability. During the development process, there are still problems such as difficulty in maintaining trajectory tracking motion accuracy and system stability under large load, strong impact and dynamic load conditions, multiple constraints leading to uneven motion trajectory and reduced trajectory tracking ability, and traditional control algorithms are currently unable to meet the requirements of high-precision trajectory tracking of robots. For this reason, a robot tracking error control method based on neural network model predictive compensation is proposed. Summary of the invention
[0007] The purpose of the present invention is to provide a robot tracking error control method based on neural network model prediction compensation to solve the problems raised in the above background technology.
[0008] To achieve the above object, the present invention provides the following technical solution: a robot tracking error control method based on neural network model prediction compensation, applying the "sequence to sequence" processing method of the recurrent neural network to the GACTRNN neural network, constructing the neural network model, which includes an encoder and a decoder, and the input sequence is , will be input into the network in time sequence and encoded into a semantic vector , semantic vector It is then used as the input of the decoder at each time step in the decoding process, thereby outputting the predicted sequence , and Represent the length of the input sequence and the output sequence respectively.
[0009] As a further preferred embodiment of the present technical solution: when processing discrete-time tasks, the neural unit is used as a discrete model, and the neuron parameter calculation and update method is as follows:
[0010]
[0011] .
[0012] As a further preferred embodiment of this technical solution: For time series input, is the time series forecast output, , At time step and The internal state of the neuron is and are the external and internal weight matrices of the neuron, respectively. is the bias, is the neuron activation function, Use the hyperbolic tangent function In order to achieve faster convergence speed.
[0013] As a further preferred embodiment of this technical solution: The time scale parameter of the GACTRNN neuron is used to indicate the update speed of the neuron's internal state and the training parameter , and ,in and are the weight matrices for the input and the neuron activity at the previous time step, is the bias of the time scale parameter. From the formula, we can see that the time series input of the neural network is At each time step, it is fed into the model and combined with the matrix Multiply, and then the time scale of the current time step To regulate, due to the parameters Is with external input Therefore, GACTRNN can regulate the update / forgetting rate of internal neurons according to the current external input to achieve a prediction effect based on the external state.
[0014] As a further preferred embodiment of the present technical solution: after completing the construction of the GACTRNN network, it is necessary to evaluate the capability of the constructed neural network model on the device, and explore the performance of the model through actual data processing tasks. The present neural network learns and generalizes the trajectory tracking error of the industrial robot.
[0015] As a further preferred embodiment of the present technical solution: in exploring the learning and generalization capabilities of the neural network model, both are based on the processing of timing error signals. According to the differences in the model sequence input mode and sequence output mode, this paper proposes two prediction modes, namely "closed-loop mode" and "open-loop mode".
[0016] As a further preferred embodiment of the present technical solution: when studying the learning ability of the model, a "closed-loop mode" is adopted, and when studying the generalization ability of the model, an "open-loop mode" is adopted.
[0017] As a further preferred embodiment of the present technical solution: in the "closed-loop mode", the latter part of the input data of the model's previous prediction and all the previous prediction results are used as input for the next prediction, and the prediction output of the model is added to the input through an active loop. The initial input sequence data needs to be provided during the first prediction of the model, and each subsequent prediction model will no longer require external data input, but will be able to continuously perform continuous predictions based on its last prediction value.
[0018] As a further preferred embodiment of the present technical solution: in the "open-loop mode", the current external signal is input into the network in real time to make a prediction. This method is the general application of neural networks, that is, a set of data is given to the trained network to obtain a set of output data. The trained network is used to process other data with different characteristics from the training samples, and error evaluation is performed to evaluate the processing ability of the trained network for other feature data. The stronger this processing ability is, the stronger the generalization ability of the model is.
[0019] As a further preferred embodiment of the present technical solution: the GACTRNN network is a recurrent neural network. When processing the error data of a time series, it is necessary to divide the input data into sliding windows of a specified length, and continuously input the windows into the neural network for iterative learning of the errors, thereby completing the model training, obtaining the corresponding prediction error results, and performing targeted compensation for the industrial robot in the Cartesian workspace according to the prediction results.
[0020] Compared with the prior art, the present invention has the following beneficial effects: The present invention constructs a neural network model to learn and generalize the trajectory tracking error of the industrial robot. When processing the error data of the time series, the input data is divided into sliding windows of specified length, which are continuously input into the neural network for iterative learning of the error, thereby completing the model training and obtaining the corresponding prediction error results. According to the prediction results, targeted compensation is performed on the industrial robot in the Cartesian workspace to maintain the motion accuracy of trajectory tracking and the stability of the system, thereby meeting the requirements of high-precision trajectory tracking of the robot. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a structural diagram of the GACTRNN neural network model; Figure 2 Schematic diagram of the structure of GACTRNN neurons; Figure 3 Schematic diagram of neural network used for industrial robot trajectory tracking error learning; Figure 4 A schematic diagram of the sequence generation of the neural network training methods "open loop mode" and "closed loop mode"; Figure 5 Schematic diagram of the method for generating continuous error prediction sequences for industrial robot errors. DETAILED DESCRIPTION
[0022] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application, and cannot be understood as limitations on the present application. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0023] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "install", "connect", "connect", and "set" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.
[0024] See also Figure 1-5 The present invention provides a technical solution: a robot tracking error control method based on neural network model prediction compensation, which applies the "sequence to sequence" processing method of the recurrent neural network to the GACTRNN neural network to construct the neural network model. The neural network model includes an encoder and a decoder. The input sequence is , will be input into the network in time sequence and encoded into a semantic vector , semantic vector It is then used as the input of the decoder at each time step in the decoding process, thereby outputting the predicted sequence , and Represent the length of the input sequence and the output sequence respectively.
[0025] In this embodiment, specifically: when processing discrete-time tasks, the neural unit is used as a discrete model, and the neuron parameter calculation and update method is as follows:
[0026]
[0027] .
[0028] In this embodiment, specifically: For time series input, is the time series forecast output, , At time step and The internal state of the neuron is and are the external and internal weight matrices of the neuron, respectively. is the bias, is the neuron activation function, Use the hyperbolic tangent function In order to achieve faster convergence speed.
[0029] In this embodiment, specifically: The time scale parameter of the GACTRNN neuron is used to indicate the update speed of the neuron's internal state and the training parameter , and ,in and are the weight matrices for the input and the neuron activity at the previous time step, is the bias of the time scale parameter. From the formula, we can see that the time series input of the neural network is At each time step, it is fed into the model and combined with the matrix Multiply, and then the time scale of the current time step To regulate.
[0030] Under the above settings, due to the parameter Is with external input Therefore, GACTRNN can regulate the update / forgetting rate of internal neurons according to the current external input to achieve a prediction effect based on the external state.
[0031] In this embodiment, specifically: after completing the construction of the GACTRNN network, it is necessary to evaluate the model capabilities of the constructed neural network on the device, explore the performance of the model through actual data processing tasks, and learn and generalize the trajectory tracking error of the industrial robot.
[0032] In this embodiment, specifically: in exploring the learning and generalization capabilities of the neural network model, based on the processing of timing error signals, according to the differences in model sequence input and sequence output methods, this paper proposes two prediction modes, namely "closed-loop mode" and "open-loop mode".
[0033] In this embodiment, specifically: when studying the learning ability of the model, the "closed-loop mode" is adopted, and when studying the generalization ability of the model, the "open-loop mode" is adopted.
[0034] In this embodiment, specifically: in the "closed loop mode", the latter part of the input data of the model's previous prediction and all the previous prediction results are used as input for the next prediction, and the model's prediction output is added to the input through an active loop.
[0035] Under the above settings, the initial input sequence data needs to be provided when the model makes its first prediction. After that, the model will no longer require external data input for each prediction, but will be able to make continuous predictions based on its last prediction value.
[0036] In this embodiment, specifically: in the "open-loop mode", the current external signal is input into the network in real time to make a prediction. This method is the general application of neural networks, that is, a set of data is given to the trained network to obtain a set of output data. The trained network is used to process other data with different characteristics from the training samples, and error evaluation is performed to evaluate the trained network's ability to process other feature data. The stronger this processing capability is, the stronger the generalization ability of the model is.
[0037] In this embodiment, specifically: the GACTRNN network is a recurrent neural network. When processing the error data of a time series, it is necessary to divide the input data into sliding windows of a specified length, and continuously input them into the neural network for iterative learning of the errors, thereby completing the model training and obtaining the corresponding prediction error results. Based on the prediction results, targeted compensation is performed on the Cartesian workspace of the industrial robot to maintain the motion accuracy of trajectory tracking and the stability of the system, thereby meeting the requirements of high-precision trajectory tracking of the robot.
[0038] The working principle of the present invention is: by building a neural network model to learn and generalize the trajectory tracking error of the industrial robot, when studying the learning ability of the model, the latter part of the input data of the model's last prediction and all the prediction results of the last prediction are used as the input of the next prediction, and the prediction output of the model is added to the input through an active loop. The initial input sequence data needs to be provided during the first prediction of the model. After that, each prediction model will no longer require external data input, but can continuously make continuous predictions based on its last prediction value. When studying the generalization ability of the model, the current external signal is input into the network in real time to make a prediction, that is, to give A set of data is sent into the trained network to obtain a set of output data. The trained network is used to process other data with different characteristics from the training samples and error evaluation is performed to evaluate the processing ability of the trained network for other feature data. When processing the error data of the time series, the input data needs to be divided into sliding windows of specified length and continuously input into the neural network for iterative learning of the error, so as to complete the model training and obtain the corresponding prediction error results. According to the prediction results, targeted compensation is performed on the Cartesian workspace of the industrial robot to maintain the motion accuracy and system stability of trajectory tracking, so as to meet the requirements of high-precision trajectory tracking of the robot.
[0039] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. Robot tracking error control method based on neural network model prediction compensation, Features: The "sequence to sequence" processing method of the recurrent neural network is applied to the GACTRNN neural network to build the neural network model. The neural network model includes an encoder and a decoder. The input sequence is , will be input into the network in time sequence and encoded into a semantic vector , semantic vector It is then used as the input of the decoder at each time step in the decoding process, thereby outputting the predicted sequence , and Represent the length of the input sequence and the output sequence respectively.
2. The robot tracking error control method based on neural network model prediction compensation according to claim 1, Features: When processing discrete-time tasks, the neural unit is used as a discrete model, and its neuron parameters are calculated and updated as follows: 。 3. The robot tracking error control method based on neural network model prediction compensation according to claim 2, Features: For time series input, is the time series forecast output, , At time step and The internal state of the neuron is and are the external and internal weight matrices of the neuron, respectively. is the bias, is the activation function of the neuron, Use the hyperbolic tangent function .
4. The robot tracking error control method based on neural network model prediction compensation according to claim 3, Features: The time scale parameter of the GACTRNN neuron is used to indicate the update speed of the neuron's internal state and the training parameter , and ,in and are the weight matrices for the input and the neuron activity at the previous time step, is the bias of the time scale parameter, the time series input of the neural network At each time step, it is fed into the model and combined with the matrix Multiply, and then the time scale of the current time step To regulate.
5. The robot tracking error control method based on neural network model prediction compensation according to claim 4, Features: After completing the construction of the GACTRNN network, the capability of the constructed neural network model is evaluated on the device, the performance of the model is explored through actual data processing tasks, and the trajectory tracking error of the industrial robot is learned and generalized.
6. The robot tracking error control method based on neural network model prediction compensation according to claim 5, Features: In exploring the learning and generalization capabilities of the neural network model, based on the processing of timing error signals, two prediction modes, namely "closed-loop mode" and "open-loop mode", were proposed according to the differences in model sequence input and sequence output methods.
7. The robot tracking error control method based on neural network model prediction compensation according to claim 6, Features: in, When studying the model's learning ability, the "closed-loop mode" is adopted, and when studying the model's generalization ability, the "open-loop mode" is adopted.
8. The robot tracking error control method based on neural network model prediction compensation according to claim 7, Features: In "closed-loop mode", the latter part of the input data of the model's previous prediction and the entire prediction results of the previous prediction are used as input for the next prediction, and the model's predicted output is added to the input through the active loop.
9. The robot tracking error control method based on neural network model prediction compensation according to claim 8, Features: In the "open-loop mode", the current external signal is input into the network in real time to make a prediction. This method is the general application of neural networks, that is, a set of data is sent to the trained network to obtain a set of output data. The trained network is used to process other data with different characteristics from the training samples, and error evaluation is performed to evaluate the trained network's ability to process other feature data.
10. The robot tracking error control method based on neural network model prediction compensation according to claim 9, Features: The GACTRNN network is a recurrent neural network. When processing the error data of time series, it is necessary to divide the input data into sliding windows of specified length and continuously input them into the neural network for iterative learning of the error, so as to complete the model training and obtain the corresponding prediction error results. Based on the prediction results, targeted compensation is performed on the Cartesian workspace of the industrial robot.
Citation Information
Patent Citations
Semi-parameterized industrial robot dynamics modeling method of convolutional neural network
CN112327630A
Robot switching type predictive control trajectory tracking method based on neural network model
CN115344047A
Multi-mode hybrid control method based on lower limb exoskeleton robot
CN115381672A
High-speed maneuvering target tracking error compensation method based on time sequence intelligent smoothing
CN116993778A
Method and a system for predicting time series values using an artificial neural network
RU2744041C1
Cited By
Flexible joint mechanical arm tracking control method and device based on adaptive neural network
CN120116212A
Robot hybrid intelligent self-adaptive control method and system adopting reinforcement learning
CN120715906A
Motion track prediction method of cross-medium aircraft and computer program product
CN121457522A
Method for predicting and compensating trajectory error of industrial robot based on non-autoregressive network
CN122378748A
Method for predicting compensation of trajectory errors of an industrial robot based on a non-autoregressive network
CN122378748B
